Deep Residual Transfer Ensemble Model for mRNA Gene-Expression-Based Breast Cancer

The last few years have witnessed exponential rise in breast cancer disease. The increasing mortality rate due to the lack of earlier diagnosis has alarmed healthcare industry to develop more efficient and scalable computer aided diagnosis (CAD) solution. Despite that the biopsy-based histopathological analysis provides easier breast cancer detection, yet the invasive treatment and iterative costs limit its scalability for the masses. Gene-expression analysis can be an alternative, where the different machine learning and artificial intelligence techniques can be developed to learn microarray RNA details for breast cancer detection. Unlike traditional methods where phenotypic changes are learnt or detected by machine learning models, the use of mRNA specific two-dimensional images can yield superior results due to increased spatial informativeness. Unfortunately, the major existing methods use convolutional neural networks (CNN) to learn features from the RNA-specific 2D images. The inability to exploit contextual details, long-term dependency, gradient vanishing and high dependence on local hierarchical features confine their efficacy in run-time environments. This research proposes deep residual transfer ensemble model for mRNA gene-expression based breast cancer detection. The proposed method at first transformed mRNA one-dimensional sequences into 2D images. Executing standard pre-processing tasks, the synthetic minority over-sampling technique (SMOTE) was applied to alleviate any possible class-imbalance problem. Subsequently, ResNet101 and AlexNet deep networks were applied distinctly to extract respective features, which were later fused to yield composite ensemble feature. Here, ResNet101 avoided gradient vanishing, while AlexNet provided 4096-dimensional features to train the ensemble-of-ensemble (E2E) classifier for consensus-based breast cancer detection. The E2E ensemble learning method used bagging, AdaBoost, Random Forest, Extra Tree Classifier and XGBoost algorithms as base classifier to perform maximum voting-based prediction. Being consensus-driven solution, it improved reliability of breast cancer prediction results. Simulation results confirmed its robustness in terms of superior accuracy (99.91%), precision (99.65%), recall (99.81%) and F-Measure (99.72%) over other methods, signifying its optimality towards real-time and scalable breast cancer diagnosis.